Performance

Quake3 Arena

Keeping in mind how greatly the performance drops with tri-linear filtering enabled, we decided to run the tests in Quake3 Arena with both: bi-linear and tri-linear filtering:

Bilinear filtering helped Kyro II to catch up with GeForce2 GTS in 32bit color mode, but in 16bit mode, where the memory bus was loaded not so heavily, Kyro II fell behind GeForce2 GTS.

When we enabled tri-linear texture filtering, Kyro II showed the lowest performance in 16bit color mode and in 32bit mode it slightly outperformed GeForce2 MX 400.

Hercules 3D Prophet 4500 didn't experience a striking twist in performance as it passed from 16bit to 32bit color mode. It lets us assume that at high resolutions Kyro II's performance is limited by insufficient fillrate.

3DMark 2001

At this stage we expanded the number of racers and added Creative 3D Blaster TNT2 Ultra graphics card. Its core and memory were clocked at the same 175MHz as those of Kyro II. Like Kyro II, NVIDIA TNT2 Ultra chip has two pixel pipelines (with one texturing unit each) and has no hardware T&L unit. So, it may come in handy to compare the performance of Kyro II and TNT2 Ultra working at the same frequencies and to find out the weak points of Kyro II.

Low Detail Mode:

Game 1 - Car Chase:

Game 2 - Dragothic:

Game 3 - Lobby:

High Detail Mode:

Game 1 - Car Chase:

Game 2 - Dragothic:

Game 3 - Lobby:

Kyro II has demonstrated astonishing results: in all the gaming tests except "Dragothic" in low-detail mode Kyro II goes abreast with GeForce2 MX, but as the resolution grows in 32bit mode Kyro II surpasses GeForce2 GTS, which appeared possible due to the advantages of the chip tile architecture, no doubt.

But what do we see in high detail mode in "Dragothic" benchmark? In this case the performance of Kyro II depends neither on the resolution, nor on the color depth, but something restricts it heavily. In high detail mode in "Car Chase" Kyro II is even slower than TNT2 Ultra!

Well, we'll make an attempt to spell it out.

The main difference between "High Detail" and "Low Detail" modes is the number of polygons that doubles in all the gaming tests. The amount of textures doubles too and some objects are overlaid by more texture layers.

If the local graphics memory were unable to store all the textures, they'd be transferred via AGP. But for these tests the top amount of textures per frame is 36MB (for 32bit color mode of "Car Chase"). It's by all means not enough to make the 64MB Kyro II transfer the textures via AGP bus. Moreover, if it were so, the results in 16bit and 32bit modes would differ greatly, since 16bit textures occupy two times less space. However, we didn't notice any striking difference: Kyro II gave nearly the same performance in high-detail mode regardless of the color depth.

You won't scare Kyro II by the necessity to lay a large number of textures: it can lay up to 8 textures per pass. Besides the test results would be greatly determined by the resolution, which is also not the case of Kyro II.

There are only two parameters left, which could have an influence like that on the performance: polygons transfer rate via AGP bus and polygons processing speed provided by the chip's core. Let's turn to "Dragothic" test, which doesn't load the CPU with extra calculations having nothing to do with the scene creation. In high-detail mode an average scene comprises about 100 thousand polygons. Kyro II makes circa 16 frames per second. It means that in a second the chip transfers approximately 100,000 x 16 polygons to the graphics card. If we suppose that the polygons were transferred as stripes, it will take about 50Bytes to describe each triangle, i.e. the graphics card receives an average of 100,000 x 50 x 16Bytes per second (it makes 80 million Bytes of polygon data). On the other hand, Wcpuid utility indicated that Hercules 3D Prophet 4500 based on Kyro II worked in AGP 2x mode with enabled SBA (side band addressing). In this case, AGP bus bandwidth is 66,000,000 x 4 x 2 = 532 million Bytes per second. As you see, the figures are too different even if we consider all the assumptions and inaccuracies. So, the conclusion is that AGP bus bandwidth doesn't restrain the performance of Kyro II.

The only factor left is the speed of polygons processing in the core. Don't miss the fact that tile architecture makes Kyro II chip sort the polygons before building a scene. So, this specific trait acts as a restrictive factor for Kyro II performance when the chip undergoes the tests with a big number of polygons.

We used a synthetic benchmark called "High Polygon Count" from 3DMark2001 to illustrate this idea. The Kyro II was clocked at 175MHz and 185MHz:

By the way, the dinosaurs from this test consist of so many polygons that now and then the lists of tile polygons in the core get overfilled and arouse some artifacts:

High Polygon Count by Kyro II

Enlarged fragment

Unreal Tournament

In Unreal Tournament Hercules 3D Prophet 4500 boasted higher performance than NVIDIA GeForce2 GTS (even in 16bit mode, not to mention 32-bit). This occurred due to relatively few polygons per scene in Unreal Tournament combined with pretty high Overdraw value. Furthermore, we enabled bi-linear filtering giving Kyro II a chance to beat its competitors.

Serious Sam

In this test Kyro II gets close to GeForce2 MX only at the highest resolutions in 32bit color mode. If you wish to make the system show polygons (the level is built of) and stripes (specially arranged groups of polygons the models are built of), you need to enter the following commands: wld_bShowPolygons=1 and mdl_bShowStripes=1:

DemoSP03 Scene


DemoSP03 Scene with wld_bShowPolygons=1 and mdl_bShowStripes=1

According to these screenshots, we can conclude that the Overdraw is quite low while the number of polygons is quite high, therefore, Kyro II "feels ill-at-ease" as it has to wait for the CPU to provide it with all the polygons of the scene and then to sort them before building the scene. Worse still, the absence of T&L unit also limits the performance of Hercules 3D Prophet 4500 in Serious Sam.

Giants

In this game Kyro II doesn't boast high performance either, because the number of polygons is big enough. Then, Giants is designed for graphics cards with T&L unit and all the graphics cards we selected for our investigation (except Hercules 3D Prophet 4500) have hardware T&L unit.

2D Image Quality

Hercules 3D Prophet 4500 offers high quality 2D graphics: we increased the resolution up to 1600x1200 and noticed no blurring effect. To assess 2D graphics quality of Hercules 3D Prophet 4500 we used two displays - ViewSonic P775 and Samsung Syncmaster 900 IFT.

Conclusion

Well, we've just got acquainted with an offspring of PowerVR and SÒMicroelectronics which provoked keen interest of Hercules Company. This company made a brave move having launched a Kyro II based graphics card named 3D Prophet 4500. Most probably, Hercules will release the whole set of Kyro and Kyro II based graphics cards. We expect to see cards with a TV-Out and DVI interface.

The shift from 0.25micron to 0.18micron manufacturing technology allowed the developers to increase the clock frequency, but all Kyro's weak points were nevertheless inherited in full by Kyro II.

Kyro II differs favorably from low-end and mainstream NVIDIA chips thanks to its beautiful performance in games with a not very big number of polygons involved. However in the newest games developed for DirectX7 and up, that is for graphics cards with hardware T&L unit, Kyro II is most likely to fail.

Sad but true: Kyro II is fairly considered a little bit outdated. That is why it's likely to be welcomed only by the fans of older games such as Unreal, Unreal Tournament, Quake3 and the games based on these engines. Of course, if you are a tile architecture admirer, are fond of some exotic products and dislike NVIDIA, you'll find Hercules 3D Prophet 4500 a good choice for relatively small money. Besides, Kyro/Kyro II driver developers tried to foresee nearly everything to make Kyro and Kyro II work properly with those applications and games, which are not specially designed for tile based accelerators. Most people, however, strongly associate "NVIDIA" with "high quality and performance" and hence do not regard Kyro II as a worthy competitor to NVIDIA stuff.

We believe that the best way to raise Kyro II's rating would be to announce graphics cards by Hercules with 32MB graphics memory. This shouldn't tell too negatively on the performance of Kyro II graphics cards, but is sure to make them still cheaper: these cards will be bundled with low-cost SDRAM chips with 5ns access time. PowerColor may decide to launch a Kyro II graphics card too, but considering NVIDIA's pressure upon other manufacturers we wouldn't claim that it will really happen.

As a final word, we would like to remind you of the highs and lows of Hercules 3D Prophet 4500 and Kyro II:

Highs:

• High performance in games with big Overdraw;
• Outstanding graphics quality in 16bit color mode;
• The ability to lay up to 8 textures per clock, EMBM and Dot3 support;
• No artifacts and compatibility with almost all the popular games;
• Good 2D quality;
• Quality mounting;
• Low price.

Lows:

• A grave performance drop when tri-linear and anisotropic filtering are enabled;
• Low performance in games with a lot of polygons (caused by tile rendering peculiarities);
• No T&L unit;
• Some bugs in drivers.